Chaitanya Mitash

I am a Ph.D Candidate in the Computer Science department at Rutgers, co-advised by professors Abdeslam Boularias and Kostas Bekris . My research interests lie in the field of Robotics, Computer Vision and Machine Learning. Specifically, I am working on problems related to perception for Robotic Manipulation.
During my PhD, I did internships with Microsoft Hololens, Redmond, WA where I worked on semantic segmentation for mixed reality applications and at Amazon Robotics, North Reading, MA where I worked on task-driven perception for manipulation.
Before graduate school, I worked as a software engineer in the R&D divisions of Samsung Electronics and Tejas Networks in the embedded software domain. I received my B.E. degree in Computer Science from Birla Institute of Technology, Mesra in 2012.

Email  /  CV  /  Google Scholar  /  GitHub  /  LinkedIn  /  Thesis  /  Thesis Presentation

An integrated perception and manipulation planning pipeline with a dynamic object representation, demonstrating high success rate and efficiency in solving pick-and-constrained placement of previously unseen objects.

Learning in simulation to predict residual poses for object models in RGB-D sequences.

A perception and motion planning framework that considers the uncertainty in object poses to generate picking plans that minimizes the chance of collision and maximizes the likelihood of success.

Simultaneous pose estimation of robot hand and the hand-held object via 3d pointset registration while considering physical consistency.

Presents a set of interpretive principles for how a robotic arm can use pointing actions to communicate task information to people. The evaluation distinguishes two classes of pointing actions that arise in pick-and- place tasks: referential pointing (identifying objects) and locating pointing (identifying locations).

Introduces key machine learning operations for joint 6D pose estimation of multiple instances of objects in challenging scenarios by learning over just simulated data.

Introduces manipulation primitives (toppling, adaptive pushing and fine-correction) to achieve robust bin packing with a simple vacuum-based end-effector. A robotic system is develeoped and large-scale experiments are performed to demonstrate its efficacy.

Sim2real adaptation of object detector via weakly labeled images and utilization of activation maps for 6d object pose estimation.

A soft segmentation output can be used to guide sampling-based pointset registration to retrieve object poses.

Integrates the self-supervised learning via physics simulation with online scene-level reasoning.

A combinatorial search algorithm to generate, evaluate and select physically-consistent scene hypothesis from independent object detections.

Physically-realistic simulation allows better training for object detection in robotics setup and can bootstrap a self-learning process to automatically collect and label real images and improve the detector over time.